Two tokens looks like extra work until you see what it buys: the AI Guard credential lives in Pangea Vault with automatic rotation, and the config only carries a fetcher token and an item ID. The audit log is the other half — every screening decision is recorded, so a blocked message can be explained after the fact rather than just disappearing. It only wraps stdio servers, so remote ones need `mcp-remote` in between.
A proxy that sits between an MCP client and any MCP server, routing all tool input and output through Pangea AI Guard. You keep the server you were already using; the proxy is what the client launches, with the original command as its argument.
- Prompt injections blocked, including ones wrapped in a riddle
- Malicious links, IPs and domains caught, using CrowdStrike, DomainTools and WhoisXML threat intelligence
- 50 types of confidential information and PII detected
- 10 content filters covering toxicity, self harm, violence and topic-based filtering, across 104 spoken languages
- Your AI Guard token held in Pangea Vault with automatic rotation, rather than sitting in your client config
- Every AI Guard request and its detection results written to your Secure Audit Log, giving an immutable trail for audits and incident response
Node.js v22.15.0 or greater, and two Pangea API tokens: one with access to AI Guard, stored in Pangea Vault, and one with Vault access used to fetch the first at runtime. In the client config, wrap the original command — `npx -y @pangeacyber/mcp-proxy -- <original command and args>` — and set `PANGEA_VAULT_TOKEN` to the Vault API token and `PANGEA_VAULT_ITEM_ID` to the Vault item holding the AI Guard token. Remote servers over HTTP or SSE need `mcp-remote` first to turn them into stdio. Binary name `mcp-proxy`.
One command plus a key — npx -y @pangeacyber/mcp-proxy, then supply credentials